Respiratory diseases are common on poultry farms during spring and autumn. Due to the high density of the farm environment, an epidemic can spread very quickly and cause large-scale biological infections. Therefore, developing software capable of monitoring or providing early warnings of respiratory disease in chickens is very important because it helps prevent the spread of diseases and enhances the health of the chickens. This study proposes an acoustic detection system for chicken coughing (ASCT-CC) designed for real-world poultry farming environments. This system is based on an improved audio spectrogram transformer (AST) architecture, using a hybrid convolutional-transformer backbone network that replaces the global attention mechanism with local multi-head attention. This architecture effectively improves the model’s ability to capture crucial local acoustic information and increases its robustness against noise at a lower computation cost. Besides, the study constructs a two-branch co-learning structure, adopts focal loss as an auxiliary strategy to reduce sample bias, and combines these with the connectionist temporal classification (CTC) decoder to accurately identify and temporally localize the cough event. To meet practical application requirements, the system is deployed with low latency on edge computing devices using TensorRT acceleration and INT8 quantization technology. Experiments demonstrated that our model achieves an mAP of 92.86% during training and reaches an identification rate of 92.11% on the independent test set, with an inference time of only around 200 ms. This system provides 24 h real-time monitoring and multi-level early warning capabilities, offering effective technical support for the early detection and intelligent control of respiratory diseases in poultry.
Cai et al. (2026) studied this question.